EXTRAPOLATION WITH MACHINE LEARNING BASED EARLY-STAGE ENERGY PREDICTION MODELS

Abstract

Machine learning (ML) approaches are necessary to quickly predict energy performance at early stages. Although ML predictions are primarily useful for interpolation and many applications, early design stages require flexibility due to uncertain parameter values and vaguely defined building forms. We tested the ability of two ML approaches, artificial neural network (ANN) and convolutional neural network (CNN), to extrapolate for early-stage energy predictions. While ANN uses numerical parameters to represent building shape, CNN uses convoluted layers to learn geometrical representations. The generalisation of both networks deteriorates on extrapolated datasets with the CNN having a better performance than the ANN.